Deep Fish

Build, run & share AI media workflows by @noahsolomon

[noah @noahsolomon]

it’s honestly kind of wild how fun it is to just iterate on a character identity with ai

watching something evolve over time from a rough idea into something that actually feels coherent is a totally new kind of creative loop

without tools like fal this just wasn't possible. and even wasn't possible until around a year ago in a reasonable amount of time / effort

I never really bought the whole ai is going to replace hollywood / artists thing that feels like such a narrow read of what’s actually going on

what does feel real is this shift in how we express and transmit ideas. gen ai isn’t about making content faster it’s about making ideas more sticky, more personal, and more direct

the ai slop shit is just a phase like most things but underneath that, something way bigger is happening to me

learning feels different. creative expression feels different. the barrier between thinking something and showing someone that thing is collapsing into the same point

that’s the interesting part to me. expanding who gets to create and how ideas move between people

it feels a new medium entirely

a more personal way of communicating

[noah @noahsolomon]

If you want to build a node editor with react flow you should check out the Deep Fish repository.

Left is showcased open source react flow component vs. deep fish's on the right.

https://t.co/HKUkwy7uHK

[Deep Fish @Deepfishlol]

@noahsolomon 🥹🐟

[noah @noahsolomon]

Deep Fish is now Open Source! @deepfishlol

My goal was to build the web version of comfyui that had the extensibility of touch designer.

I ended short so graph execution is not recursive and remains client directed.

Go drop it a star and if anyone needs help plz dm :D

[noah @noahsolomon]

going to open source @deepfishlol this weekend

[noah @noahsolomon]

This is what i've been working on for the PAST 2 MONTHS!!

Fal Sandbox where you can run models on the same input and see the (shareable) generations in one view!

https://t.co/SMY80TI1rV

Hope you guys like it (another 20 trillion to fal) 🥹🫶

[noah @noahsolomon]

Sunsetting development on @deepfishlol.

When a side project begins to feel more like a chore than an expression of curiosity I think that's when it's the best time to move on.

I've began to work on a heart rate / hrv wristband monitor last week and want to move more in this direction (intersection of health monitoring hardware and data analysis) for my side projects through the next year.

If u want to help and live in sf lmk :D

[Lisbon AI @lisbonai_]

We’ve all seen AI prototypes that wow in a demo… but never make it to real products.

Now, with AI, it's even easier to have amazing demos, but the problem still persists.

@noahsolomon turns those AI ‘toys’ into tools teams actually use. At @fal, he builds fast and ships tools bridging experiments to production.

At LisbonAI (LXAI), you’ll be able to talk with Noah about the craft of productizing AI.

Noah is a builder by heart:

→ Created brainrot.js - a text-to-video experiment that took over timelines
→ Built Infinite Kanvas - a multiplayer gen-art canvas powered by Fal
→ Shiped Deepfish - workflow + inference infrastructure for complex AI pipelines
→ Hacked PPO Bunny - browser-based RL on ONNX with a 3D UI
→ Prototyped Pillzumi - narrative agents with memory and personality
→ Released code-first tools like Deep Star and Compositional Token Adaptation
→ Once built a multiplayer 3D platformer in the browser - now wrangles workflows, weird UIs, and wandering gradients

Builders should be among builders to learn with each other.

👉 Join us at LisbonAI: https://t.co/k8y7R5JKoh

[Deep Fish @Deepfishlol]

@noahsolomon deep fish hat

[noah @noahsolomon]

I love my stupid pfp

[noah @noahsolomon]

Real time persistence maxxing 🐟⋆｡𖦹°🫧˖

there's like 0 local state for workflows not directly anchored to server state now (much needed improvement!).

(for workflows ✅ for flows 🔜)

changelog 08/31 - @deepfishlol :D

[noah @noahsolomon]

User downloaded .deepfish files for workflows are going to be encoded in characters like ⋅˚ ଳ ˚ 𓆝⋆｡˚ 𓇼 ⋅˚ ଳ ˚ 𓆝⋆｡˚ 𓇼

so when u open the file it’s just an ocean in text I think it’s pretty cool 😸

[noah @noahsolomon]

Character design in vintage anime style using Dreamina 3.1

[noah @noahsolomon]

It’s always such a warm feeling when u see an app u have rank for new keywords on google

[noah @noahsolomon]

Going to give a shot at making a short anime episode 100% AI this weekend that doesn't feel AI.

If you'd like to try too here's the way of doing so with consistent style, consistent voices, limited artifacts in video scenes, and captions that i'm thinking of from models that are all on @fal (minus RVC right now):

Each scene needs to be grounded with a start and end frame. We can get the first frame from Nano Banana from google. However for text to image, it's not the best model for aesthetics. So I'll use Dreamina 3.1 to get the initial character designs and environments and overall style/aesthetic from text prompts and from then on use Nano banana to make variations of those characters and environments.

For video generation we could either do a start+end frame video interpolation approach but the problem with that for dialog heavy scenes is it's not entirely automatable to guarantee generated audio and interpolated video match pacing or even length actually. And even if they match on both fronts, then you need to use an external lipsync model which always has artifacting (and is trained for realism so idek how it would do for like a dragon talking for example, I imagine pretty bad).

The best approach for the video side I think is start frame into veo 3 is the best approach. Then with that audio+video output, transcribe who is speaking when in the video with the eleven labs transcription segment api and then isolate the vocals during those time segments, and run the vocals through a rvc model for the voice you want to use for that character (for consistent voice) and combine the rvc outputted vocal with the vocal removed audio and replace the audio for that scene with the voice remapped audio.

And for captioning, there's a transcription + video editing model on fal to add captions to the video.

[Deep Fish @Deepfishlol]

Converting this red dragon into a diamond dragon with nano banana

[noah @noahsolomon]

Qwen Image

[noah @noahsolomon]

My 3 favorite text to image models for vintage anime aesthetic are Qwen Image, Dreamina and Wan 2.2

I made a flow for my 3 favorite anime models you can go try for free with different characters and backgrounds.

link below

[noah @noahsolomon]

Nano Banana is really awesome but I'm not impressed with it's aesthetics for text to image.

I still think the best text to image model for aesthetics in general is wan 2.2 right now

[Deep Fish @Deepfishlol]

@0xGoodfuture https://t.co/dnO0q42od4

[Deep Fish @Deepfishlol]

Go try nano banana, the best image editing model for free right now!

Here are some flows we've built with it too 🧵

[Deep Fish @Deepfishlol]

@0xGoodfuture it's already here